New Research arrived
For quite a while progress of modern AI has been closely tied to the model itself. New capabilities are created through training, encoded into parameters and improved through larger architectures, better data and more compute. Even continual learning — the field concerned with enabling AI to acquire new knowledge without losing what it has already learned — has largely followed this model-centric view.
A recent paper suggests that this boundary is beginning to move.
In Continual Learning in Transition, Hou et al. describe a shift from parameter-centric learning toward system-level adaptation. Rather than treating model parameters as the only place where capability can evolve, the authors broaden the perspective to include external structures such as memory, skill libraries and interaction protocols. Together with developments in inference-time and on-policy learning, these mechanisms extend adaptation beyond the static parameter space of an individual model.
This may appear to be a technical evolution within continual learning. We believe its implications are more fundamental.
If meaningful capability can increasingly develop outside the model itself, then the model may no longer be the natural boundary of an intelligent system.
Where does learning actually happen?
Classical continual learning has focused heavily on the challenge of changing a model without destroying previously acquired knowledge. This idea remains important, but the emerging system-level view expands the question from how a model changes to where learning can occur at all.
Hou et al. organize this transition around three dimensions: when learning takes place, how adaptation happens and where evolving capability resides. In this framework, learning can extend from pre-training and post-training into inference time, beyond conventional gradient-based updates, and from internal parameters into external structures surrounding the model.
The consequence is subtle but important. An AI system can develop useful capabilities without requiring every development to be internalized in the weights of a neural network.
Memory can persist beyond an individual interaction. Skills can be accumulated and reused. Protocols can shape future behavior. Experience can influence subsequent actions after deployment.
The evolutionary boundary of the system becomes larger than the parameter space of any individual model.
This changes the role of the model. It remains an extraordinarily powerful component, but it does not necessarily have to contain the complete evolving capability of the system around it.
From better Models to developing Systems
This distinction matters because much of the current AI race is still organized around a model-centric assumption — with significant implications for energy consumption and a financial bet, which might not pay off.
If we want more capable AI, we build a more capable model. If reasoning is insufficient, we improve the model. If knowledge is missing, we expand training or context. If behavior needs to change, we fine-tune.
System-level adaptation introduces a different path.
Capability may increasingly emerge from the relationship between multiple persistent elements of an AI system rather than from improvements to a single component in isolation.
That does not mean models become less important. It means the architecture around them becomes more important.
The relevant question therefore begins to change from:
How can we build a model that continues to learn?
to:
How can we build an intelligent system that continues to develop?
For ONONO, this distinction is fundamental.
The Model as a Component of Intelligence
We have previously argued that memory alone is not sufficient for persistent machine understanding.
Retrieving previous information is different from understanding how that information relates to a changing context. New evidence can alter the meaning of earlier events. Assumptions can become outdated. Intent can evolve. Different interpretations can remain plausible at the same time.
This is why we introduced the concept of a Mind Ontology: a persistent cognitive structure in which memory is part of a larger system for maintaining context, relationships, uncertainty and evolving meaning over time.
The emerging transition in continual-learning research approaches the problem from another direction.
If adaptation itself increasingly moves beyond the parameters of a model and into the surrounding system, then the architecture responsible for organizing that system becomes increasingly consequential.
In other words:
The model may become a component of intelligence rather than its container.
This is a substantially different way of thinking about AI.
It does not require one model to contain everything an intelligent system knows, learns or becomes capable of doing. Instead, it opens the possibility of persistent systems whose capabilities and understanding can develop across time.
Beyond Continual Learning
Continual Learning in Transition does not propose a Digital Mind, nor does it validate ONONO's architecture. But the direction it describes is significant: adaptation is beginning to move beyond the model and into the surrounding system.
For ONONO, this points to the larger architectural question behind Progressive AI: how can an intelligent system develop persistently as information, context and reality change?
We believe that as such systems gain continuity and the ability to refine their understanding over time, another abstraction becomes increasingly relevant:
the Digital Mind.
From the model to the system — and potentially, from the system to the mind.
Related ONONO Research
The Scientific Foundations of ONONO: Why Organizations Need a Digital Mind
A broader examination of the limitations of model-centric AI and the scientific foundations behind Progressive AI.
Memory Is Not Enough: Why AI Needs a Mind Ontology
Why persistent memory alone cannot provide evolving context and understanding — and why a higher-level cognitive structure is required.
Reference
Zhiyan Hou, Dan Zhang, Tao Feng, Liyuan Wang, Wei Li, Xiangzhao Hao, Hongyan An, Junfeng Fang, Haokai Ma, Zhaohui Xu, Haiyun Guo, Jinqiao Wang & Tat-Seng Chua. Continual Learning in Transition . arXiv:2608.06216. August 6, 2026.